Engineering Structures 178 (2019) 603–615

Contents lists available at ScienceDirect

Engineering Structures
journal homepage: www.elsevier.com/locate/engstruct

Recurrent neural network model with Bayesian training and mutual T

information for response prediction of large buildings
Carlos A. Perez-Ramireza, Juan P. Amezquita-Sancheza, , Martin Valtierra-Rodrigueza,
⁎

Hojjat Adelib, , Aurelio Dominguez-Gonzaleza, Rene J. Romero-Troncosoa

⁎⁎

a
Faculty of Engineering, Autonomous University of Queretaro, Campus San Juan del Rio, Rio Moctezuma 249, Col. San Cayetano, 76805 San Juan del Rio, Mexico
b
Department of Biomedical Informatics, Neurology, and Neuroscience, Columbus, OH 43220, USA

ARTICLE INFO ABSTRACT

Keywords: An accurate response prediction model is of great importance in various applications such as damage detection,
Highrise building structures structural health monitoring, and vibration control. Development of such a methodology for large civil structures
Vibrations is challenging because of their size and complicated behavior and noise-contaminated, nonlinear, and nonsta-
Structural system identification tionary nature of the signals. In addition, the prediction model must have a low computational burden for real-
Empirical mode decomposition
time applications. In this article, a new methodology and a nonlinear autoregressive exogenous model (NARX)-
Non-linear autoregressive exogenous model
based recurrent neural network (NN) model is presented for accurate response prediction of large structures. The
Neural networks
methodology is based on adroit integration of three concepts: a recent signal processing concept, empirical mode
decomposition (EMD), mutual information (MI) index from the information theory, and a probabilistic Bayesian-
based training algorithm. The EMD method is used to remove the noise in the measured signals. An MI index is
proposed to determine the optimum number of neurons in the hidden layer of the NN model with the goal of
reducing the computational requirements without affecting its performance. Finally, Bayesian regularization
(BR) is proposed to train the optimized NN model. The effectiveness of the proposed methodology is assessed by
predicting the structural response of a 1:20-scaled 38-story highrise building structure subjected to seismic
excitations and ambient vibrations, and a five-story steel frame subjected to different levels of the Kobe
earthquake.

1. Introduction feature extraction and classification for SHM are discussed by Amez-
quita-Sanchez and Adeli [3]. Among recent works, Zhi et al. [63] pre-
In the past few decades, design and construction of highrise and sent an inverse identification method for estimation of structural re-
super-highrise buildings have surged as an economical alternative to sponses of highrise building structures under wind loads. Shan et al.
provide both working and living spaces in large cities of the world [57]. [48] examine identification of building structures under seismic loading
In this regard, it is necessary to develop algorithms that can con- using performance and stabilizing objectives. Kim et al. [28] consider
tinuously assess the building condition in order to avoid potentially modal identification of highrise building structures based on ortho-
severe economic and human life losses due to extreme events such as gonality of filtered response vectors. Yin et al. [61] describe an entropy-
earthquakes and high winds. This technology is referred to as structural based optimal sensor placement strategy for model identification of
health monitoring (SHM) [4,49,54]. A key step in SHM is often struc- periodically articulated bolted structures. Oh et al. [38] discuss modal
tural system identification (SSI), a process for estimating the structural response-based visual system identification and model updating ap-
response of structures from a set of input-output measurements pro- proaches for building structures. Park et al. [40] take up vision-based
duced by dynamic signals [50]. SSI finds application in damage de- natural frequency identification using laser speckle imaging and par-
tection [4,7,51], modal parameters identification [39], and vibration allel processing.
control [6,58]. A review of structural system identification up to 2012 The estimation of the structural response under unknown condi-
is presented by Sirca and Adeli [50] followed by a review of SHM by tions, however, is a challenging task especially considering the struc-
Qarib and Adeli [43]. On the other hand, the techniques used for tural response signals are noise-corrupted, non-linear, and non-

⁎
Corresponding author.
⁎⁎
Corresponding author.
E-mail addresses: [email protected] (J.P. Amezquita-Sanchez), [email protected] (H. Adeli).

https://doi.org/10.1016/j.engstruct.2018.10.065
Received 7 March 2018; Received in revised form 22 September 2018
Available online 26 October 2018
0141-0296/ © 2018 Elsevier Ltd. All rights reserved.
C.A. Perez-Ramirez et al. Engineering Structures 178 (2019) 603–615

Fig. 1. Flowchart of the proposed methodology: (a) Main steps and (b) Neuron selection using MI.

stationary. The higher the accuracy of the SSI model, the higher the influenced by the choice of the mother wavelet and level decomposition
accuracy of the structural response estimation. In the literature, para- as well as the noise contained in the signal [4]. Jiang et al. [24] note
metric and non-parametric approaches have been developed to provide that the efficacy of typical artificial neural network (NN)-based models
different solutions for the SSI problem [2]. Parametric approaches are is reduced when signals have very complex input patterns. To lessen
used mainly to model structural responses with linear or time-invariant this issue, they introduced the concept of stochastic neural network
features which allow calculating some physical parameters of the (SNN) for SSI, where the typical activation functions used in traditional
structures such as stiffness, natural frequencies, and damping ratios. NN are replaced with a stochastic function such as a Gaussian function.
These methods may work for small structures in the elastic range; but, In this regard, the combination of effective denoising strategies as well
they often fail to estimate the nonlinear behavior of large civil struc- as the introduction of the probabilistic concepts into the NN [62] model
tures under unknown and varied dynamic loading accurately [2]. Fur- will result in more robust methodologies, capable of dealing with the
thermore, they are susceptible to the noise contained in the measured inherent noise present in the acquired samples. Further, a probabilistic
signal, resulting in errors in the estimated model. In order to overcome approach for the training scheme can improve the model generalization
the aforementioned drawbacks, non-parametric approaches have been capabilities leading to improved prediction abilities. Moreover, an en-
developed with promising results, representing a superior alternative hanced model will be computationally more efficient, a desirable fea-
for structural response prediction [5,24]. ture for real-life implementation of the model in large civil structures
Jiang and Adeli [21] introduced the concept of wavelet neural such as highrise building structures.
networks (WNN) in vibration control of structures as an effective This paper presents a new recurrent NN-based model for non-
method for approximating and predicting the nonlinear behavior of parametric identification of large civil structures subjected to dynamic
large structures. Subsequently, WNN was used by Lee and Shin [30] in loadings, where a recurrent topology allows improving the prediction
constrained approximate optimization and Jiang et al. [24], Mitchell capabilities of the model developed by Adeli and Jiang [2]. The pro-
et al. [35], Zhou et al. [64], and Khalid et al. [27] in structural system posed methodology is based on the adroit integration of three concepts,
identification. However, two issues must be addressed for a WNN model a recent signal processing concept, that is, empirical mode decom-
to yield accurate results. First, the number of neurons in the hidden position (EMD), mutual information (MI) index from the information
layer of the network must be optimized to maximize accuracy and theory, and a probabilistic Bayesian-based training algorithm [36]. The
minimize the computational burden. Second, the performance of WNN EMD method is used to remove the noise in the measured signals [44].
is degraded when the data to be processed are embedded in high-level An MI index is proposed to determine the optimum number of neurons
noise, thus requiring an effective signal denoising step. To deal with in the hidden layer of the NN with the goal of reducing the computa-
these issues for large civil structures, Adeli and Jiang [2] proposed a tional requirements without affecting its performance. Finally, Bayesian
multi-paradigm methodology through adroit integration of three dif- regularization (BR) is proposed to train the optimized NN model. Even
ferent computing paradigms: chaos theory, fuzzy logic, and the WNN when the signals used to construct the model are denoised in the de-
concepts. The methodology was validated using different levels of the noising stage, it is not possible to fully assess whether there is still any
Kobe earthquake. They report their methodology can predict the re- noise remaining in the denoised data, thus introducing a component of
sponse of a 1/2-scaled five-story steel frame with a maximum error of uncertainty in the measured data. The utilization of the BR allows in-
3%. corporating this uncertainty indirectly [23] as the resulting model will
Despite promising results reported in the literature, further research be of a stochastic nature. Moreover, this training scheme improves the
on the two aforementioned aspects can lead to additional improve- error minimization, thus ensuring good generalization capabilities for
ments in both the prediction error and the computational efficiency of the proposed model [26]. The usefulness and effectiveness of the pro-
the obtained structural response prediction models. Therefore, alter- posed methodology are assessed by predicting the structural response of
native approaches should be investigated for optimizing the number of a 1:20-scaled 38-story highrise building structure subjected to seismic
neurons in the hidden layer to allow selecting only the hidden layer excitations and ambient vibrations, and a five-story steel frame sub-
neurons that contain relevant information [13] and for denoising the jected to different levels of the Kobe earthquake.
data, especially since the Wavelet Transform (WT) results can be

604
C.A. Perez-Ramirez et al. Engineering Structures 178 (2019) 603–615

Fig. 2. (a) Schematic diagram of an NARX-NN model and (b) Bipolar sigmoid function.

2. Proposed methodology Unlike wavelet thresholding, the EMD method requires neither the
utilization of a predefined function to perform the decomposition nor
Fig. 1a shows a macro-flowchart of the proposed methodology the setting of a finite number of decomposition levels, allowing its ef-
consisting of three main steps. The EMD method is used in Step 1, as a fective utilization for denoising.
filtering approach to denoise the time-series signal, mainly the high- EMD method extracts in a sequential form the time scales, i.e., IMFs
frequency noise. In Step 2, a nonlinear autoregressive exogenous presented in the signal, starting from the components that have a finer
(NARX) neural network (NN) model (NARX-NN) coupled with a Baye- temporal scale (high-frequency modes) to the ones that have a wider
sian training scheme is used to create a baseline model to predict the scale (or low-frequency modes). Hence, the sum of all the obtained
structural responses of civil structures. Finally, in Step 3, the baseline modes or IMFs, i.e., the reconstruction of the input signal, produces a
model is computationally improved by selecting only the neurons in the complete match with the original signal. It has been noted that for
hidden layer that contribute to the estimation of the desired output signals corrupted by broadband noise, the first IMF has the lowest
significantly using the MI method. The resulting model is retrained in signal-to-noise ratio; in other words, the first IMF has most of the noise
order to ensure the best possible prediction results. The following sub- contained in the signal [8,11,14]. Thus, elimination of the first IMF
sections describe the proposed methodology in detail. allows denoising the signal without modifying its frequency content. An
analytical expression for the noise contained in the signal cannot be
developed due to the nature of EMD method [11]. While some noise
2.1. Step 1 EMD-based filtering approach may still remain in the signal after removing the first IMF, removing
additional IMFs could eliminate useful information. As such, it is not
The denoising of non-linear time-series signals requires the utiliza- recommended to remove additional IMFs because it can affect the fre-
tion of time-frequency algorithms capable of detecting sharp edges and quency content adversely and impact the model prediction capability
short-duration impulses [11]. In this regard, methods based on wavelet negatively [23]. The BR algorithm can handle any remaining noise
thresholding, such as soft and hard thresholding as well as universal without affecting the quality and efficacy of the model prediction
thresholding schemes have been used [10,25,52,60]. These methods are capabilities.
based on the assumption that the signal magnitude is higher than the For vibration signals obtained from civil structures, the signals of
noise magnitude. In other words, the signal-to-noise ratio (SNR) has a interest are usually concentrated in the low-frequency region (e.g., last
large value, thus, the wavelet coefficients whose values are lower than a IMFs), especially for real-life buildings [17,24,32,56]. Therefore, the
predefined threshold value can be set to zero without altering the actual EMD method is used in this work to decompose the signal into its IMFs.
signal where the threshold value is related to the noise amplitude The first IMF is removed as it contains most of the noise [11], and the
[11,29]. However, many real-life signals, and in particular vibration remaining IMFs are used to construct a new signal. For a complete
signals from large civil structures, are often embedded into high-level explanation of the EMD algorithm see Huang et al. [18] and Amezquita-
noise. This condition causes a decrement in the SNR value, making the Sanchez and Adeli [4].
selection of the appropriate thresholding level a difficult task. In ad-
dition, a proper selection of the mother wavelet and the number of
decomposition levels is necessary in order to ensure the best possible 2.2. Step 2 Constructing and training the model
results [31,47].
In this research, an alternative adaptive method, EMD, is explored After denoising the input signal using the EMD method, a recurrent
for analyzing time-series signals with noisy, non-linear and non-sta- NN is constructed and employed to predict the response of the civil
tionary features. In this approach, the signal is decomposed into a finite structure under dynamic excitations such as earthquakes and ambient
number of band-limited signals known as intrinsic mode functions vibrations. A NARX-NN approach is described conceptually by the fol-
(IMFs), according to information contained in the analyzed signal. lowing equation [42]:

605
C.A. Perez-Ramirez et al. Engineering Structures 178 (2019) 603–615

n
y (t ) = f [y (t 1), …, y (t n y ), u (t 1), …, u (t nu )] + e (t ) (1)
e= [y (i ) y (i )]T [y (i) y (i)]
(6)
where u(·), y(·), e(t), nu and ny, are the model input and output, the noise i=1

term, and the maximum lags for the input and output, respectively. where y and y are the vectors of the actual and predicted values, re-
Noise is assumed to be a distributed random variable with a normal spectively. In order to mitigate the overfitting problem found in some
distribution. Fig. 2(a) shows a schematic representation of the NARX- training algorithms such as the LMBP method, the BR approach as-
NN architecture. This architecture has three layers: input-, hidden-, and sumes that the function has a certain amount of smoothness, that is,
output-layers [42]. The input layer is denoted by the time delay units (z- when the connection weights in the network have a small value, the
1
) which are the past samples required by the model to predict the output of the network will be smooth [33]. Using Eq. (6), the BR
response. The hidden layer is composed of a finite number of neurons to scheme uses a regularization factor in the following form (F) [15,16]:
map the relationship between the set of input data and the corre-
F = e + ws (7)
sponding output data. The output layer multiplies the output of each
where ws = k = 1 [w (k )]2 ; K indicates the total number of weights
K
hidden layer neuron by its corresponding weight and adds them up to
estimate the model output. Researchers have noted if ny and nu are set employed by the model, and α and β are the parameters to be optimized
to have the same value, the network offers the best compromise be- using an optimization algorithm. The regularization is a process used to
tween the prediction error and the computational effort required [19]. restrict the magnitude of the weights of the NARX-NN model. A smaller
Hung et al. [19] and Jiang and Adeli [21] found that a value of 2 for value will be less prone to overfitting the data used to train the model
both ny and nu provides the best results. Hence, this value is used in this [16]. It should be pointed out that when α ≪ β the training algorithm
research. will focus on reducing the error between the actual and predicted
The output of the hidden layer, yp(t), is estimated using: outputs. In contrast, when α ≫ β the training algorithm will focus on
reducing the weight size. This reduction will help the model generate a
K K smooth output, which will further reduce the noise contained in the
yp (t ) = f1 IWi ui + IWo yj + b1 training signal [16]. The goal of a regularization scheme is to select the
i=1 j =1 (2) values of α and β properly. The BR approach does not require a priori
knowledge [33]; α and β are considered random variables. The optimal
where f1 is the bipolar sigmoid function, shown in Fig. 2(b), IWi and IWo
weights must minimize the regularized objective function, Eq. (7). In
are the connection weights for u and y, respectively; K is the number of
this regard, the terms of Eq. (7) are employed to develop the following
neurons in the hidden layer, and b1 is the bias term used for the hidden
conditional probability function [33]:
layer. The bias term allows capturing the linear behavior of a signal on
top of the nonlinear behavior [22]. The bipolar sigmoid function (f1) is P (D , , M ) =
P (D w, , M ) p (w , M )
P (w D, , , M )
used because it provides improved convergence for the training algo- 1 1
e ( e) e ( ws )
rithm compared with the standard sigmoid function [9]. The bipolar zD ( ) zW ( )
=
sigmoid function is defined as follows [59]: 1
zF ( , )
e (F )

zF ( , )
2 =
f1 = 1 zD ( ) z W ( ) (8)
1+e 2k (3)
where D represents the training data set, M is the NARX-NN model
where k is the value of the inside of the parenthesis in Eq. (2). f1 is used used, P(D|w, β, M) is the conditional probability function of the data
to compute the model output, y (t ) , as follows: occurrence given the weight vector w, P(w|α, M) is the prior weight
K
density function, which represents the knowledge of the weights. In
y (t ) = OWi f1i + b2 general terms, the variance of the analyzed signal will indicate the
i=1 (4) magnitude of the weights. Before starting the training procedure, the
magnitude of the noise depends on the noise present in the signal. The
OWi represents the connection weights of links connecting the nodes in noisier the signal, the smaller the values of the weights to smoothen the
the hidden layer to the output node, and b2 is the bias term. It should be predicted signal. In contrast, the cleaner the signal, the larger the values
noted that a training algorithm must be used to find the values of IWi, of the weights.
IWo, OWi, b1, and b2. For the sake of clarity of presentation, all the P(w|D, α, β, M) is the density function for w, and zD(β), zW(α), and
weights used by the network are placed in a weight vector w, defined as zF(α,β) are constants related to the number of weights and biases used
follows: by the network. In particular, the probability density function P(D|w, β,
M) is estimated assuming that the errors in the measured signal, i.e.
w = [IWi IWo OWi] (5)
measurement and quantization noise as well as the prediction errors,
Training algorithms such as the classic Levenberg-Marquardt are independent and follow a Gaussian distribution. On the other hand,
backpropagation (LMBP) and gradient-based methods [1] are often the functions for P(w|α, M) and P(w|D, α, β, M) are estimated assuming
employed to train the NN-based model due to their ability to deliver than the weights have a zero-mean condition with a Gaussian dis-
fast results when the size of the network is relatively small [20]. These tribution. The equations for these probability density functions are as
methods, however, are prone to overfitting problems, decreasing the follows [15]:
network’s ability to correctly predict the model output when new input 1 e
P (D w , , M ) = e
values are presented, that is, its generalization ability [34]. Moreover, zD ( ) (9a)
these methods have difficulties to capture the stochastic features that
real-life structures generally present [24]. MacKay [33] proposed a 1 ws
p (w , M ) = e
Bayesian training algorithm. The utilization of such a stochastic-based zW ( ) (9b)
training algorithm allows dealing with the stochastic nature of signals. n
2
Mahapatra and Sood [34] noted that this scheme has an improved zD ( ) =
prediction performance. In this research, a Bayesian regularization (BR) (9c)
is proposed for training the NARX-NN model. N
In general terms, the main objective of a training process is to zW ( ) =
2

minimize the sum of the squared errors defined as follows [33]: (9d)

606
C.A. Perez-Ramirez et al. Engineering Structures 178 (2019) 603–615

1 which is the sum of the terms of the diagonal of H. The Levenberg-

zF ( , ) ( (2 ) ) N
2
det(H MP )
e F(w MP)
(9e)
Marquardt algorithm is used as an efficient numerical method for the
computation of H and J [26,34,38]. It is not used as a training algo-
The BR algorithm is used to obtain the optimal weights that mini- rithm in this research because it can lead to overfitting problems [34].
mize the error between the measured outputs and the computed outputs Fig. 3a and b shows the steps used to construct and train a NARX-
of the model through the regularization factor (F) and the probability NN-based model, respectively. The delays for the input (nu) and output
function defined by Eq. (8). The probability function consists of three (ny) are set to 2 because this value leads to the best possible results
constants: zD, zW, and ZF (Eqs. (9a)–(9e)) where the last constant re- [19,21]. At the beginning, the number of the neurons in the hidden
quires the calculation of the Hessian matrix (H) which allows knowing layer is initialized arbitrarily, keeping in mind that the maximum
if w produces the minimum error for the predicted values. H can be number of neurons should be less than the number of samples used for
calculated using H = 2 F but this approach results in high computa- training [16].
tional costs [15,16]. In order to reduce the computational cost, Foresee Once the non-optimized NARX-NN model is configured in Step 1, it
and Hagan [15] introduced an alternative approximate expression for H is trained using the BR algorithm in Step 2. Fig. 3b shows the steps of
as follows: the BR-based training algorithm described as follows:
H 2 JT J + 2 I (10)
(1) Initialize α = 0 and β = 1. These weights can also be initialized
where J is the Jacobian matrix of the training set errors, i.e., [15], the randomly with a uniform distribution with values within the −1 to
exponent T denotes the transpose of J, and I is the identity matrix. To 1 interval [16].
compute the optimal values of α and β, the derivatives of the natural (2) Compute the regularization factor, F, defined by Eq. (7).
logarithm of Eq. (7) with respect to α and β are set equal to zero [16]. (3) Compute the Hessian matrix, H, and the effective number of para-
On the other hand, γ, the effective number of parameters, an integer meters, γ, employing Eqs. (10) and (13).
number, is obtained by using the eigenvalues of H [16]: (4) Calculate the new values of the parameters α and β in the reg-
ularization factor equation using Eqs. (11) and (12). Iterate be-
MP = tween steps 2 and 4 until convergence is achieved within a given
2ws (w MP) (11)
tolerance.
MP n
=
2e (w MP) (12)
2.3. Stage 3 Improving the computational efficiency of the proposed NARX-
2 MP NN model
=N
tr (H MP ) (13)
The purpose of this step is to select the neurons that contribute the
where the exponent MP denotes the values of α and β that minimize Eq. most for the accurate estimation of the desired response while avoiding
(7), n is the total number of weights, and tr denotes the trace of matrix, the overfitting problem and reducing the computational resource

Fig. 3. Flowchart for: (a) constructing a NARX-NN model, (b) BR training algorithm, and (c) algorithm for selection of the neurons in the hidden layer.

607
C.A. Perez-Ramirez et al. Engineering Structures 178 (2019) 603–615

requirement. It is noted that an overfitted model produces a degraded An important aspect of the proposed methodology is the selection of
response when new (unseen) inputs/examples are applied to the model neurons that contribute significantly to the estimation of the desired
with reduced prediction capabilities. response. Fig. 3c shows the algorithm for the selection of the neurons in
MI allows capturing and measuring the dependence between two the hidden layer in the NARX-NN model. In order to determine the
variables. In this research, MI is used to measure the dependence be- appropriate number of neurons, a thresholding scheme is developed
tween the output of a particular neuron and the desired response so that based on the estimated MI values. This threshold is selected so that only
the redundant neurons, those that do not have a relationship with the neurons above it contribute significantly to the desired response and are
output, can be safely removed. The dependence between the output of a selected in the NARX-NN model, and the remaining neurons are dis-
neuron and the desired response is determined by: carded. The resulting number of neurons is thus considered to be ap-
propriate for the predicting the response.
p (x , y )
I (x ; y ) = p (x , y ) log The proposed threshold (t) is defined as the mean (µ) of the esti-
y Y x X
p (x ) p (y ) (14) mated MI values for all neurons in the hidden layer plus the standard
deviation (σ) of the MI values divided by three:
where I(x;y) is the MI between the variables x and y, p(x) and p(y) are
the probability functions of x and y, respectively, and p(x,y) is the joint
t=µ+
probability density function of x and y. These functions are defined as 3 (18)
follows:
The effectiveness of the proposed threshold scheme will be verified
2
1 (
1 x µx
) via numerical experimentation in the next section.
p (x ) = e 2 x
x 2 (15)

1 y µy
2 3. Validation of proposed model
1 2
p (y ) = e y
y 2 (16) 3.1. 38-Story highrise building
where σx, μx, σy, and μy are the variance and mean for x and y, re-
spectively. The computation of p(x,y) requires the knowledge of the 3.1.1. Description of the verification example
conditional probability of x given y, which in practical terms is a dif- In order to evaluate the performance of the proposed model, the
ficult task since an analytic function cannot be properly described [12]. experimentally measured data from a 1:20 scaled model of a high-rise
For this reason, the Parzen method is employed to estimate p(x,y) [53]. residential building shown in Fig. 4 subjected to seismic excitations and
In this article, x denotes the output produced by the th-neuron in the ambient vibrations obtained by Ni et al. [37] are used. The structure
hidden layer and y is the measured response of the structure. MI is used consists of three podium floors (floors F1 to F3), one transfer floor (floor
to find how much of x is contained in y. It is postulated that the neurons F4), 34 typical floors (floors F5 to F38) and three top floors (floors F39
whose MI values exceed a threshold value have a significant contribu- to F41) (see Fig. 5). It is instrumented using accelerometers, which are
tion to the desired output [41]. Most existing thresholding methods do
not use information about the particular process under consideration. In
contrast, the proposed methodology uses statistical properties of the
signal to determine the threshold value. This ensures that the threshold
value is estimated for a specific case. This threshold is determined using
an expression that combines the mean and standard deviation. Using
this scheme, an appropriate threshold value is estimated in order to
discover the neurons that have the strongest contribution to the pre-
dicted output. Further, this scheme relies only on statistical properties
rather than assumptions that can change when the algorithm is applied
to other scenarios. The value of MI varies from 1 for a neuron with a
high contribution to the desired output to 0 for a neuron with no de-
pendence on the desired output that can be safely removed. The re-
sulting model is re-trained in order to ensure the best possible results.
To evaluate the model obtained, the relative root mean square
(RMSE) error between the measured output and that predicted by the
NARX-NN model is used as a performance indicator [2,19]:
N
k=1
(yp (k ) y (k )) 2
RMSE =
N
k=1
(yp (k ) y (k ))2 (17)
where y(k) and y (k ) are the measured output and the mean of the
predicted outputs, respectively. A small RMSE value indicates the
model has an excellent prediction performance.
Processing speed and computation time is an important considera-
tion in developing a prediction model for real-time application in large
real-life structures. One way to improve the computational efficiency of
the algorithm is to use the minimum number of neurons in the hidden
layer that can yield accurate results. In this research, this is achieved
using an MI-based algorithm allowing to select only the neurons that
contribute to the output significantly. Further, this process avoids the
use of redundant information provided by the neurons, which helps
mitigate the overfitting problem resulting in improved model predic-
tion capabilities. Fig. 4. Scaled model of the 38-story residential building.

608
C.A. Perez-Ramirez et al. Engineering Structures 178 (2019) 603–615

Fig. 5. (a) Substructure identification and (b) Sensor locations for the highrise structure shown in Fig. 4.

distributed in the 9 substructures (noted by dashed line in Fig. 5a). The responses of the floors F34 and F25, and the computed responses from
locations of the accelerometers are also noted in Fig. 5b by solid circles. the last two previous time steps from the model of the floor F30 are
The scaled model of the building was designed and dynamically used as inputs to predict the response of the floor F30.
excited at the Hong Kong Polytechnic University using a 5 × 5 m 6-DOF Following the proposed methodology, the measured responses are
shaking table. Two types of excitations were applied to the scaled first denoised using the EMD. Then, a non-optimized NARX-NN with 50
building. Following results are used to verify the proposed identifica- neurons in the hidden layer combined with the BR training scheme are
tion system: seismic excitation (Case 1) and seismic excitation followed used for estimating the response of floor F30. The initial selection of the
by low‐intensity ambient vibrations (white noise) (Case 2). In Case 1, number of the neurons in the hidden layer is a non-trivial choice;
the new computational model is tested using three different levels of however, it is recommended to start with a rather large network [16].
artificially generated seismic excitations (small-, moderate-, and large- After performing several numerical tests, it was found that the use of
level) corresponding to three different levels of damage (small, mod- more than 50 neurons in the hidden layer did not significantly change
erate, and large damage). The damage levels were characterized ac- the predicted response; thus, this number was chosen for the initial
cording to visual inspection of the cracks and their widths in the number of hidden neurons. Next, the MI method is applied to select the
structural members of the building. The structural responses for the neurons that contribute to the estimation of the predicted response
small-damage level were acquired during a 45 s period using a 100 Hz- significantly.
sampling rate resulting in 4500 samples. On the other hand, responses The proposed threshold value defined by Eq. (18) was determined
for the other two damage levels, moderate and large, were obtained based on an extensive numerical experimentation using different
during a 30 s period using the same sampling rate resulting in 3000 numbers of neurons in the hidden layer and predicting the response of
samples. floor F30 using the proposed methodology.
In Case 2, after the structure is excited with the artificially gener- Fig. 6a–c shows the values of MI for the NARX-NN model for various
ated seismic signals, white noise excitations are applied to the structure neurons in the hidden layer for 50, 30, and 10 neurons in the hidden
in order to simulate low intensity ambient vibrations and determine one layer, respectively. The threshold value obtained using Eq. (18) is
of the five states of the structure: no damage, small damage, moderate identified by a red1 dash horizontal line. The large green circles above
damage, and large damage. The structure is excited for 200 s using the the threshold line indicate the neurons that make a significant con-
same 100 Hz-sampling rate. This results in 20,000 samples. For a more tribution to the response. The three models with three different num-
detailed explanation of tests performed refer to Ni et al. [37]. bers of nodes in the hidden layer (50, 30 and 10 neurons) yield the same
Similar to previous works that have used these data [22,37] and optimal number of neurons in the hidden layer, that is 9, indicating that
with the purpose of comparing the results obtained in this research with regardless of the number of neurons in the hidden layer chosen at the
those obtained by Jiang and Adeli [22], only the structural responses in outset, the model leads to the same optimal number of nodes in the
the x-direction and measured by sensors located at F25, F30, and F34 hidden layer. Thus, it is concluded that the optimal number of neurons
floors (Fig. 5a) are used to construct and validate the model. to predict the response is equal to 9 (neurons 1, 18, 20, 23, 26, 34, 35,
41, and 45, highlighted by large circles in Fig. 6a–c).
Fig. 7a shows the predicted response of the floor F30 using the re-
3.1.2. Case 1 Seismic excitations
corded signals without denoising. Fig. 7b displays the predicted re-
In this example, the 4500 samples recorded during the small seismic
sponse using the signals denoised by the EMD method and a non-op-
excitation in the x-direction of the floors F34, F30, and F25 are used to
timized NARX-NN model. Fig. 7c presents the predicted response using
construct the proposed NARX-NN model. The data are divided into two
the signals denoised by the EMD method and the optimized NARX-NN
parts: the first 2800 samples are used for training the optimized NARX-
model. The original response is represented by a solid (blue) line, while
NN model, and the other 1700 samples are used for validation purposes.
Similar to Jiang and Adeli [22], for training the model, the acquired
acceleration responses at floors F25 and F34 are used as inputs, and the
actual acceleration responses at floor F30 as the output of the NARX-NN 1
For interpretation of color in Figs. 6, 7, 9, the reader is referred to the web
model. On the other hand, for validating the proposed model, the version of this article.

609
C.A. Perez-Ramirez et al. Engineering Structures 178 (2019) 603–615

Fig. 6. MI values for the NARX-NN model using (a) 50, (b) 30, and (c) 10 neurons in the hidden layer.

Fig. 7. Comparison of the predicted responses of floor F30 during the small seismic excitation by the NARX-NN models created with (a) noisy signal, (b) denoised
signal with non-optimized number of neurons, and (c) denoised signal with the optimized number of neurons.

the predicted response is represented by a dashed (red) line. The middle an RMSE of 0.1 indicating that the optimized model is capable of pre-
figures in Fig. 7 are the zoomed region in the range of 31–32 s. It is dicting the structural response with a high accuracy. The maximum
observed from the graphs in the middle of Fig. 7a, the case of noisy data absolute errors are reported in this article in order to compare the re-
with the non-optimized model and 7b, the case of denoised data with sults with those obtained by Jiang and Adeli [22] to be discussed
the non-optimized data, predicted responses of floor F30 using the subsequently. The estimated RMSE value for the proposed model is
NARX-NN model do not match the measured signal. Maximum absolute about 25% of the non-optimal version.
errors of 6.5 and 2 m/s2 and RMSE values of 1.5419 and 0.4327 are Once the optimized NARX-NN model with 9 neurons in the hidden
obtained for the two cases of the noisy and denoised data with the non- layer is created using the recorded accelerations during the small
optimized model, respectively. In contrast, the predicted response of seismic excitation, it is used to predict the structural response when the
floor F30 using the proposed methodology (Fig. 7c) matches the re- structure is excited with the moderate- and large-level seismic signal
ference signal closely with a maximum absolute error of 0.11 m/s2 and sets. Fig. 8a and b shows the predicted responses of floor F30 using the

610
C.A. Perez-Ramirez et al. Engineering Structures 178 (2019) 603–615

Fig. 8. Responses of floor F30 estimated by the proposed methodology subjected to (a) moderate and (b) severe seismic excitation signals.

proposed methodology subjected to moderate- and large-level seismic signals of floors F34 and F25 and the computed response in the previous
excitations, respectively. The measured and estimated responses in the two previous time steps of floor F30 are used as inputs to the model,
top figures are indistinguishable indicating the optimized NARX-NN whereas the model output is the predicted response of floor F30. The
model predicts the responses accurately. The corresponding errors are previous computed responses are used to predict the current model
displayed in the bottom figures. The maximum absolute errors for response as the model has a recurrent feature defined by Eq. (1). It
moderate- and large-level seismic excitations are 1.3 and 1.5 m/s2, re- should be pointed out that models trained using data for the healthy
spectively. The RMSE values for moderate- and large-level seismic ex- condition can be employed to detect damage in the structure
citations are 0.1789 and 0.1827, respectively, indicating that the pro- [24,45,46].
posed methodology can accurately predict the structural response when Following the same sequence as in the previous case, the acquired
the structure is subjected to any level of dynamic excitations. responses are denoised using the EMD approach. Then, the optimization
Jiang and Adeli [22] presented a multi-paradigm methodology scheme is carried out by constructing several NARX-NN models with
consisting of chaos theory/nonlinear science, soft computing techni- different numbers of neurons (50, 30, and 10 neurons). Combining the
ques of fuzzy logic and neural networks, and a signal processing tech- proposed threshold (Eq. (18)) and the MI method, the neurons that
nique, wavelets, to develop a nonlinear autoregressive moving average contribute significantly to the estimation of the predicted response are
with exogenous (NARMAX) input model for predicting the structural obtained. The optimal number of neurons was found to be 9, the same
response of large structures subjected to dynamic seismic excitations. as the previous case. This value is expected as the same structure is
They applied the model to the same example structure and same seismic being analyzed. Thus, 9 hidden nodes are used in the optimized NARX-
data used in this research. They reported maximum absolute errors of NN model to predict the structural response for three cases: small-,
0.5, 1.4, and 3.5 m/s2 for small-, moderate and large-seismic excita- moderate-, and large-damage. Fig. 9a–d shows the estimated response
tions, respectively. In contrast, the methodology proposed in this paper of floor F30 for the healthy, small-, moderate-, and large-damage con-
results in maximum absolute errors of 0.10, 1.3, and 1.5 m/s2 for small-, ditions, respectively. Again, the measured (denoted by blue dashed
moderate and large-seismic excitations, respectively. These results in- lines) and estimated responses (denoted by solid red lines) in the left
dicate an improvement in the order of 7–56% for moderate to large figures are indistinguishable indicating the optimized NARX-NN model
seismic excitations. predicts the responses accurately. The corresponding errors are dis-
played in the right figures. The maximum absolute errors of 0.15, 0.4,
3.1.3. Case of ambient vibrations 0.4, and 0.3 m/s2 and the RMSE values of 0.11, 0.13, 0.23, and 0.18 are
This example is used to demonstrate the effectiveness of the pro- obtained for the aforementioned cases, respectively. Jiang and Adeli
posed methodology under noisy conditions. For this purpose, 20,000 [22] report for their model a maximum absolute error of 0.6 m/s2, for
samples acquired for the healthy condition of the building are used to the healthy, small-, moderate-, and large-damage conditions, which is
construct the optimal NARX-NN model. Half of the samples are used to higher than those reported for the model presented in this paper. The
construct and train the model and the other half is employed to verify results demonstrate that the proposed model can deal with nonlinear
the model. For training the model, the measured responses of floors F34 and noisy signals effectively. It captures the dynamic characteristics of
and F25 are used as inputs while the response of floor F30 is used as the structure accurately making it possible to estimate the progressive
output. On the other hand, for validating the model, the measured damage conditions from the healthy state.

611
C.A. Perez-Ramirez et al. Engineering Structures 178 (2019) 603–615

Fig. 9. Estimated responses of floor F30 for (a) healthy, and (b) small, (c) moderate, and (d) large damage scenarios (left figures) and corresponding errors (right
figures).

3.2. ½-Scaled five-story steel frame structure is subjected to five intensity levels of the 1995 Kobe earth-
quake signal using a shake table: 20, 32, 40, 52, and 60% of the original
In this section, the proposed methodology is applied to a ½-scaled amplitude, denoted by Kobe20, Kobe32, Kobe40, Kobe52, and Kobe60,
five-story steel frame tested at National Center for Research on respectively. The structural responses were measured at the corners of
Earthquake Engineering in Taiwan by Hung et al. [19] (Fig. 10). The each floor during 25 s using a sampling frequency of 1 kHz resulting in
25,000 samples. For a more detailed explanation of tests performed
refer to Hung et al. [19]. Fig. 11 shows the acquired/measured re-
sponses of the five floors subjected to the Kobe20 excitation signal. The
acquired samples are divided into two parts: 10,000 samples are used to
train the model and the remaining 15,000 samples are used to validate
the model accuracy. In this example, fewer training samples were used
just to test the capability of BR to capture the structural dynamics with
fewer samples. One of the benefits of employing the BR scheme as a
training algorithm is the training data set can be smaller than that re-
quired for a traditional NN because BR has an enhanced capability to
capture the relationship between the inputs and the desired output
[16,55]. Similar to the work of Adeli and Jiang [2], the structural re-
sponses of the first and third floors are used to train the model to obtain
the response of the second floor. The model validation is done by
feeding the responses of the first and third floors and the computed
response of two previous time steps of the second floor to calculate the
response of the second floor.

Fig. 11. Measured responses of the 5-story frame under Kobe20 excitation
Fig. 10. ½-scaled five-story steel frame. signal: (a) first, (b) second, (c) third, (d) fourth, and (e) fifth floors.

612
C.A. Perez-Ramirez et al. Engineering Structures 178 (2019) 603–615

Fig. 12. Estimated responses of the second floor under 5 different excitations obtained by the proposed model trained by the Kobe20 signal for (a) Kobe20 response,
(b) Kobe32 response, (c) Kobe40 response, (d) Kobe52 response, and (e) Kobe60 response.

Following the proposed model, the responses from the first, second, value of 0.0023 for the predicted response of Kobe20 excitation source,
and third floors are denoised using the EMD-based filtering scheme. a 0.0130 value for the Kobe32 excitation source, and so on. It is ob-
Next, models with 50, 30, and 10 neurons in the hidden layer are served that, generally, the proposed model can predict the responses of
constructed to determine the most appropriate number of neurons in the structure under different levels of excitation signals accurately, ef-
the hidden layer for the best possible results. The threshold estimated fectively, and reliably because the obtained RMSE values in all the
by Eq. (18) results in 5 neurons as the optimal number of neurons to scenarios are under 0.03. Adeli and Jiang [2] report an absolute max-
construct the NARX-NN model. Using the optimized NARX-NN model, imum error of 0.04 g in the best case and 0.14 g in the worst case. The
the measured responses when the structure is excited with four signals corresponding numbers for the proposed model are 0.005 g and 0.08 g
Kobe32, Kobe40, Kobe52, and Kobe60 are used to test the performance which represent improvements of over 40% and 80%, respectively. The
of the proposed model. Fig. 12a–e shows the estimated responses of the performance improvement can be explained by a) EMD appears to be
second floor under 5 different excitations obtained from the proposed more effective than wavelet-based scheme and b) utilization of the BR
model trained by the Kobe20 signal. The predicted responses nearly training algorithm allows better handling of residual noise contained in
coincide with the original ones as observed in the zoomed versions in the signal. The results presented in Section 3.1.2 show the large effect
the range 11–12 s shown as middle figures. The maximum absolute that the noise has on the model creation, as the RMSE values for the
errors are well below 0.08 g. The RMSE values obtained for the five non-optimized version are significantly greater than the denoised and
excitation sets are 0.0023, 0.0030, 0.0052, 0.0038, and 0.0082, re- non-optimized model, as seen in Fig. 7.
spectively. An interesting observation in the proposed methodology is that the
Adeli and Jiang [2] noted that the same number of neurons in the RMSE values increase when the model is trained using stronger ex-
hidden layer can be used to create and validate their model using any of citation signals. This effect may be explained as the structure properties
the five excitations (Kobe20, Kobe32, Kobe40, Kobe52, and Kobe60) change, prediction of the previous condition becomes a more challen-
because they represent the same patterns. Hence, and for comparison ging task. Nevertheless, it can be used to predict the response under
purposes with the results presented in Adeli and Jiang [2], NARX-NN lower excitation signals safely. On the other hand, the best performance
models developed for the remaining signals have the same number of case indicates that the model trained with the weakest signal, Kobe20
neurons in the hidden layer. Table 1 summarizes the RMSE values using signal, can be used to predict the structural response under stronger
the 5 excitation signals as training sets, one at a time, for the NARX-NN excitations, as the structural properties are not changed by the action of
model created with the proposed methodology along with the RMSE the excitation source.
values using the methodology proposed by Adeli and Jiang [2]. The
rows indicate the excitation source used to train and construct the
4. Conclusions
model and the columns represent the RMSE values for the predicted
signal response. The maximum errors for each condition are boldfaced.
Accurate structural system identification of large civil structures
For instance, the Kobe20-trained NARX-NN model produces an RMSE
represents a challenging problem. In this paper, a new NARX-NN-based

613
C.A. Perez-Ramirez et al. Engineering Structures 178 (2019) 603–615

Table 1
RMSE values using the 5 excitation signals as training sets, one at a time, for the NARX-NN.
Predicted response

Training excitation Kobe20 Kobe32 Kobe40 Kobe52 Kobe60

signal
New model Adeli and New model Adeli and New model Adeli and New model Adeli and New model Adeli and Jiang
Jiang [20] Jiang [20] Jiang [20] Jiang [20] [20]

Kobe20 0.0023 0.076 0.0130 0.073 0.0209 0.083 0.0207 0.08 0.0338 0.08
Kobe32 0.0030 0.077 0.0124 0.067 0.0199 0.08 0.0200 0.075 0.0317 0.075
Kobe40 0.0052 0.097 0.0130 0.094 0.0190 0.102 0.0243 0.10 0.0331 0.09
Kobe52 0.0038 0.096 0.0125 0.09 0.0193 0.092 0.0190 0.095 0.0300 0.09
Kobe60 0.0082 0.10 0.0138 0.105 0.0198 0.103 0.0198 0.105 0.0288 0.10

methodology for nonparametric identification of large civil structures Appendix A. Supplementary material
subjected to dynamic loadings is presented. Furthermore, a new
thresholding scheme combined with the MI method is proposed to de- Supplementary data to this article can be found online at https://
termine the optimal number of neurons. In order to validate the ef- doi.org/10.1016/j.engstruct.2018.10.065.
fectiveness of the proposed NARX-NN model, two examples are pre-
sented: (1) a scaled model of high-rise residential building subjected to References
seismic excitations and ambient vibrations and (2) a scaled five-story
steel frame subjected to ambient vibrations. The effectiveness of the [1] Adeli H, Hung SL. An adaptive conjugate gradient learning algorithm for effective
proposed model is compared with the methodology proposed by Adeli training of multilayer neural networks. Appl Math Comput 1994;62(1):81–102.
[2] Adeli H, Jiang X. Dynamic Fuzzy wavelet neural network model for structural
and Jiang [2] and Jiang and Adeli [22], respectively. system identification. J Struct Eng 2006;132(1):102–11.
The results for the 38-story building show that the presented [3] Amezquita-Sanchez JP, Adeli H. Feature extraction and classification techniques for
methodology is effective for predicting the structural behavior of a health monitoring of structures. Sci Iranica - Trans A Civil Eng 2015;22(6):1931–40.
(Invited Paper).
large structure subjected to dynamic excitations, since a noticeable [4] Amezquita-Sanchez JP, Adeli H. Signal processing techniques for vibration-based
resemblance with the reference signal is obtained. These results are health monitoring of smart structures. Arch Comput Meth Eng 2016;23(1):1–15.
obtained as the strategy used for mitigating the noise contained in the [5] Amezquita-Sanchez JP, Valtierra-Rodriguez M, Aldwaik M, Adeli H.
Neurocomputing in civil infrastructure. Sci Iranica - Trans A Civil Eng
signal, an EMD-based approach, is effective for handling the high-fre- 2016;23(6):2417–28.
quency noise. Based on these results, the proposed methodology is [6] Bakule L, Rehak B, Papik M. Decentralized networked control of building structures.
capable of predicting the structural behavior using signals with low- Comput-Aided Civ Infras Eng 2016;31(11):871–86.
[7] Behmanesh I, Moaveni B, Papadimitriou C. Probabilistic damage identification of a
amplitude embedded in high-level noise as the ones measured during
designed 9-story building using modal data in the presence of modeling errors. Eng
the ambient vibrations. Struct 2017;131:542–52.
In the five-story steel frame example, which exhibits a different [8] Bekara M, van der Baan M. Random and coherent noise attenuation by empirical
mechanical behavior than the 38-story building, because of the con- mode decomposition. Geophysics 2009;74(5):89–98.
[9] Bishop CM. Neural networks for pattern recognition. 1st ed. Oxford, UK: Oxford
struction material (steel) used, the results also show that the proposed University Press; 1995.
methodology is capable of predicting the structural response accurately. [10] Boto-Giralda D, Díaz-Pernas FJ, González-Ortega D, Diez-Higuera JF, Anton-
As such, it can predict the structural response in a range of behaviors, Rodríguez M, Martínez-Zarzuela M, et al. Wavelet-based denoising for traffic vo-
lume time series forecasting with self-organizing neural networks. Comput-Aided
regardless of its mechanical properties. The methodology can use fewer Civ Infras Eng 2010;25(7):530–45.
samples for the training stage than the validation one because the BR [11] Boudraa AQ, Cexus JC. EMD-based signal filtering. IEEE Trans Instrum Meas
training algorithm is capable of capturing the relationship between the 2007;56(6):2196–202.
[12] Cover TM, Thomas JA. Elements of information theory. 2nd ed. Hoboken, NJ: John
inputs and outputs accurately using a relatively small set of training Wiley and Sons; 2006.
data. [13] Dai H, Cao Z. Wavelet support vector machine-based neural network metamodel for
In summary, the proposed methodology produces excellent results structural reliability assessment. Comput-Aided Civ Infras Eng 2017;32(4):344–57.
[14] Flandrin P, Gonçalves P, Rilling G. Detrending and denoising with empirical mode
under real-life conditions with an added advantage of having low
decompositions. In: Proceedings of the 12th European signal processing conference,
computational burden employing an optimization scheme based on the Vienna, Austria; 2004.
use of the MI index. Moreover, the uncertainty present in the measured [15] Foresee FD, Hagan MT. Gauss-Newton approximation to Bayesian regularization.
In: Proceedings of the international joint conference on neural networks; 1997. p.
signal is incorporated indirectly through the BR training algorithm.
1930–35.
[16] Hagan MT, Demuth HB, Beale MH. Neural network design. 1st ed. Boston, MA: PWS
Publishing; 1996.
Acknowledgments [17] Hazra B, Sadhu A, Roffel AJ, Narasimhan S. Hybrid time-frequency blind source
separation towards ambient system identification of structures. Comput-Aided Civ
Infras Eng 2012;27(5):314–32.
The authors are grateful to Prof. Ni of the Hong Kong Polytechnic [18] Huang NE, Shen Z, Long SR, Wu MC, Shin HH, Zheng Q, et al. The empirical mode
University for providing the data for the 38-story example and Profs. decomposition and the Hilbert spectrum for nonlinear and non-stationary time
series analysis. Proc Roy Soc A: Math Phys Eng Sci 1971;1998(454):903–95.
S.H. Hung and C.S. Huang of the Taiwan’s National Center for Research [19] Hung SH, Huang CS, Wen CM, Hsu YC. Nonparametric identification of a building
on Earthquake Engineering for providing the data of the Five-Story structure from experimental data using wavelet neural network. Comput-Aided Civ
Steel Frame. Infras Eng 2003;18(5):356–68.
[20] Ilonen J, Kamarainen JK, Lampinen J. Differential evolution training algorithm for
feed-forward neural networks. Neural Proc Lett 2003;17:93–105.
[21] Jiang X, Adeli H. Dynamic wavelet neural network for nonlinear identification of
Funding highrise buildings. Comput-Aided Civ Infras Eng 2005;20(5):316–30.
[22] Jiang X, Adeli H. Pseudospectra, MUSIC, and dynamic wavelet neural network for
damage detection of highrise buildings. Int J Numer Methods Eng 2007;71:606–29.
This work was partially supported by the Mexican Council on [23] Jiang X, Mahadevan S, Adeli H. Bayesian wavelet packet denoising for structural
Science and Technology (CONACyT) under the scholarship 289377 and system identification. Struct Control Health Monit 2007;14:333–56.
[24] Jiang X, Mahadevan S, Yuan Y. Fuzzy stochastic neural network model for struc-
the SEP-CONACyT CB-2015/254697 project.
tural system identification. Mech Sys Signal Proc 2017;82:394–411.

614
C.A. Perez-Ramirez et al. Engineering Structures 178 (2019) 603–615

[25] Katicha SW, Flintsch G, Bryce J. Wavelet denoising of TSD deflection slope mea- exponentially damped signals for health monitoring of smart structures. Smart
surements for improved pavement structural evaluation. Comput-Aided Civ Infras Mater Struct 2015;24. 125040 (14 pp).
Eng 2010;29(6):399–415. [45] Rafiei MH, Adeli H. A novel machine learning based algorithm to detect damage in
[26] Kayri M. Predictive abilities of Bayesian regularization and Levenberg–Marquardt highrise building structures. Struct Des Tall Spec Build 2017;26(18). e1400 (11pp).
algorithms in artificial neural networks: a comparative empirical study on social [46] Rafiei MH, Adeli H. A novel unsupervised deep learning model for global and local
data. Math Comput Appl 2016;21(2). 20(11 pp). health condition assessment of structures. Eng Struct 2018;156(1):598–607.
[27] Khalid M, Yusof R, Joshani M, Selamat H, Joshani M. Non-linear identification of a [47] Reda Taha MM, Noureldin A, Lucero JL, Baca TJ. Wavelet transform for structural
magneto-rheological damper based on dynamic neural networks. Comput-Aided Civ health monitoring: a compendium of uses and features. Struct Health Monit
Infras Eng 2014;29(3):221–33. 2006;5(3):267–95.
[28] Kim D, Oh BK, Park HS, Shim HB, Kim J. Modal identification for high-rise building [48] Shan J, Ouyang Y, Yuan HW, Shi W. Seismic data driven identification of linear
structures using orthogonality of filtered response vectors. Comput-Aided Civ Infras physical models for building structures using performance and stabilizing objec-
Eng 2017;32(12):1064–84. tives. Comput-Aided Civ Infras Eng 2016;31(11):846–70.
[29] Kopsinis Y, McLaughlin S. Development of EMD-based denoising methods inspired [49] Shan J, Shi W, Lu X. Model reference health monitoring of hysteretic building
by wavelet thresholding. IEEE Trans Signal Process 2009;57(4):1351–62. structure using acceleration measurement with test validation. Comput-Aided Civ
[30] Lee J, Shin KH. A conservative method of wavelet neural network based meta- Infras Eng 2016;31(6):449–64.
modeling in constrained approximate optimization. Comput Struct [50] Sirca Jr. GF, Adeli H. System identification in structural engineering. Scientia
2011;89(1):109–26. Iranica 2012;19(6):1355–64.
[31] Lei Y, Lin J, He Z, Zuo MJ. A review on empirical mode decomposition in fault [51] Story BA, Fry GT. A structural impairment detection system using competitive ar-
diagnosis of rotating machinery. Mech Sys Signal Proc 2013;35:108–26. rays of artificial neural networks. Comput-Aided Civ Infras Eng 2014;29(3):180–90.
[32] Li Z, Park HS, Adeli H. New method for modal identification of super high-rise [52] Su H, Chen J, Wen Z, Wang F. Wavelet-fractal diagnosis model and its criterion for
building structures using discretized synchrosqueezed wavelet and Hilbert trans- concrete dam crack status. Trans Institute Meas Control. DOI: 10.1177/
forms. Struct Des Tall Spec Build 2017;26(3). e1312 (16 pp). 0142331217693076.
[33] MacKay DJC. Bayesian interpolation. Neural Comput 1992;4(3):415–47. [53] Tsanas A, Little MA, McSharry PE, Ramig LO. Nonlinear speech analysis algorithms
[34] Mahapatra SS, Sood AK. Bayesian regularization-based Levenberg–Marquardt mapped to a standard metric achieve clinically useful quantification of average
neural model combined with BFOA for improving surface finish of FDM processed Parkinson’s disease symptom severity. J Royal Soc Interface 2011;8:842–55.
part. Int J Adv Manuf Technol 2012;60(9):1223–35. [54] Tsogka C, Daskalakis E, Comanducci G, Ubertini F. The stretching method for vi-
[35] Mitchell R, Kim Y, El-Korchi T. System identification of smart structures using a bration-based structural health monitoring of civil structures. Comput-Aided Civ
wavelet neuro-fuzzy model. Smart Mater Struct 2012;21. 115009 (12 pp). Infras Eng 2017;32(4):288–303.
[36] Nashnush E, Vadera S. Learning cost-sensitive Bayesian networks via direct and in- [55] Ticknor JL. A Bayesian regularized artificial neural network for stock market
direct methods. Integr Comput-Aided Eng 2017;24(1):17–26. forecasting. Expert Syst Appl 2013;40(14):5501–6.
[37] Ni YQ, Zhou XT, Ko JM. Experimental investigation of seismic damage identifica- [56] Ubertini F, Gentile C, Materazzi AL. Automated modal identification in operational
tion using PCA-compressed frequency response functions and neural networks. J conditions and its application to bridges. Eng. Struct. 2013;46:264–78.
Sound Vib 2006;290(1–2):242–63. [57] Wang N, Adeli H. Sustainable building design. J Civ Eng Manage 2014;20(1):1–10.
[38] Oh BK, Kim D, Park HS. Modal response-based visual system identification and [58] Wang N, Adeli H. Self-constructing wavelet neural network algorithm for non-linear
model updating methods for building structures. Comput-Aided Civ Infras Eng control of large structures. Eng Appl Artificial Int 2015;41:249–58.
2017;32(1):34–56. [59] Yan L, Elgamal A, Cottrell GW. Substructure vibration NARX neural network ap-
[39] Osornio-Rios RA, Amezquita-Sanchez JP, Romero-Troncoso RJ, Garcia-Perez A. proach for statistical damage inference. J Eng Mech 2013;139(6):737–47.
MUSIC-ANN analysis for locating structural damages in a truss-type structure by [60] Yi TH, Li HN, Zhao XY. Noise smoothing for structural vibration test signals using
means of vibrations. Comput-Aided Civ Infras Eng 2012;27(9):687–98. an improved wavelet thresholding technique. Sensors 2010;12(8):11205–20.
[40] Park K, Torbol M, Kim S. Vision-based natural frequency identification using laser [61] Yin T, Yuen KV, Lam HF, Zhu HP. Entropy-based optimal sensor placement for
speckle imaging and parallel computing. Comput-Aided Civ Infras Eng model identification of periodically articulated structures endowed with bolted
2018;33(1):51–63. joints. Comput-Aided Civ Infras Eng 2017;32(12):1007–24.
[41] Pascoal C, Oliveira MR, Pacheco A, Valadas R. Theoretical evaluation of feature [62] Zeinalia Y, Story B. Competitive probabilistic neural network. Integr Comput-Aided
selection methods based on mutual information. Neurocomputing Eng 2017;24(2):105–18.
2017;226:168–81. [63] Zhi L, Li QS, Fang M. Identification of wind loads and estimation of structural re-
[42] Pisoni E, Farina M, Carnevale C, Piroddi L. Forecasting peak air pollution levels sponses of super-tall buildings by an inverse method. Comput-Aided Civ Infras Eng
using NARX models. Eng App Artif Intelligence 2009;22(4–5):593–602. 2016;31(12):966–82.
[43] Qarib H, Adeli H. Recent advances in health monitoring of civil structures. Sci [64] Zhou LR, Yan GR, Ou JP. Response surface method based on radial basis functions
Iranica - Trans A Civil Eng 2014;21(6):1733–42. for modeling large-scale structures in model updating. Comput-Aided Civ Infras Eng
[44] Qarib H, Adeli H. A new adaptive algorithm for automated feature extraction in 2013;28(3):210–26.

615